Hello everybody,
this is a new channel to broadcast frequent updates about SAI progress and training experiments.

These four nets were trained with KLE loss instead of MSE loss

This one instead is a "null control" where no loss term for the value head was included. Neither MSE, nor KLE but only the policy loss term. Apparently there is no significant statistical difference in a single train experiment

This is point 6/8 of a factorial experiment in 20x256 training. We are changing rate, window size, mse loss weight and steps. To assess network strength we selected a subpanel of 6 very different nets. More info when the experiment will be finished.

This is point 7/8. Next and last one in three days 😵

KLE experiment goes on, as one generation did not show measurable differences with MSE. These four nets were trained starting from the best KLE network from previous generation. We will continue in this way to see if after 5 or 10 generations we can spot some differences. Main training will ignore these networks until the end of the experiment.

Third generation of KLE nets

Fourth generation of KLE nets

Last point of the 20x256 factorial experiment

With 150 games per network, the standard error on each of these points is about 4%, which is quite large

These and all other factorial experiment matches' type renamed to "experiment"

Update on 20x256 factorial experiment after game number increase

KLE experiment 5th generation

Purple points (and lines) are KLE nets (and chosen ones). The ratings may well be inflated, in particular the last one.

This is the average performances of latest nets (one in four) against panel nets. (Redrawn with a better scale.)

7th KLE generation: d16e was selected

9th KLE generation: dbb1 was selected

10th generation: 8047 was selected

11th generation: 50d5 was selected. Subpanel matches in progress

Maybe we can insist with KLE networks a little bit more...

KLE 12th generation. d1c6 selected. We will go light with matches in KLE experiment from now on.

KLE 14th generation. c901 selected.

Again a good performance of KLE nets

KLE 16th generation. 5228 was selected

KLE 17th generation. 7776 was selected

KLE 18th generation. 983b was selected

Seems like we better consider a transition from MSE to KLE training. We will think how to do this in the best and most smooth way. SAI doesn't like discontinuities.

Added one point trained with both KLE and MSE equal to zero (only policy loss), just to check that the value head is actually being used in training currently 😅

Graph about the policy weight of best 5 first moves and the alpha of blank board (aka expected fair komi)[right axis]

Performance of last 124 generations VS panel (one point every 4 gens)

In the last three generation the promoted network was always a KLE network!

We already reported that the training D (aka 20x256 networks) failed to produce good networks. One possible reason was that we copied the training rate of previous trainings and maybe this was not right for the larger nets.
In particular, we copied the sudden drop in training rate that was *necessary* for training A (aka 6x128) —because the rate was wrong before— and still *worked* for training B (9x192) and C (12x256).
Reconsidering things a bit, we thought that maybe one problem was that this dropped the training rate too much when games were still pretty low level, and hence after that networks could find difficult to learn high-level features.
Hence we started a new experiment: redo training D copying AlphaZero training rates.
This required some maths to do the conversions, but basically at the beginning the suggested rate (0.02) was half way of our sudden drop during training A.
So we thought of doing the following: retrain D from the beginning with 0.02 until recent games, then rewind to say the last million games and train with 0.002, then rewind to the games above 10100 Elo and train with 0.0002.
We were yesterday at the end of the first swipe with high rate. Last three networks, corresponding to generations 507, 508 and 509, were trained with lower rate, to prepare before the rewind. The very last network was g1fd-792f and we decided to test it just to see how bad it was.
We did not expect this net to be any good, but we had at least to check if it was abysmal before proceeding.
We were really surprised to find that it is instead pretty good and very promising.
This may be an effect of the rate drop in nets 507-509, but it is encouraging either way.

This match is between the very last D network of the first swipe (rate 0.02, generation 544) VS the official C net of generation 543. It is confirmed that these D networks are neither very strong nor very weak. We will retrain the last 3 nets with rates 0.0093, 0.0043, 0.002, test again and start the second swipe

Sorry for the recent lack of updates. We are on the third and last swipe of D (20x256) networks, and maybe we have something

Just to be sure... we are going to test 6fd3aa against the subpanel now

Subpanel results confirm that this network is strong

ok, still preliminary, but...

😳

against LZ networks: some progress but we are still far

These are three networks of the recent D train. Their strength is confirmed.

Hello! Apparently the strength of subsequent nets of the third swipe is slowly decreasing.

Updated graph. Overfitting in the final part of the third swipe is even worse than in the second swipe. In agreement, the corresponding net of generation 528 had a terrible performance.

The test of D networks (20x256) from the recent three-swipe experiment is here.
In purple nets from the third swipe of generations from 432 to 512. In blue nets from the second swipe of earlier generations. In red, pointed by the arrow, the last net from the first swipe.

We continued the third swipe with the most recent generations and the overfitting ends, after a while.

Hello! We completed the first swipe of the above training procedure on 12x256 structure.

First network tested, compared and promoted!

The usual training over the newly promoted 186f82 network was shocking!

So, happy new year! And now I will explain the recent exploit